In a Data Science Connect webinar, Snowflake's Sam Mittal and Monte Carlo's Ethan Post explained why AI agents silently degrade after launch and how to catch it. They walked through Cortex Agents' trust features, a four-layer framework for diagnosing bad answers, a live stale-data demo and six early warning signals.
Key takeaways
- Agents degrade quietly. Mittal said production AI that works on Monday can silently degrade by Thursday when data and models drift.
- Users find most issues. In a Monte Carlo study Post cited, users found agent problems more than half the time and only 8% were caught proactively.
- Check four layers. Post's trust framework moves from output to behavior to performance to the data context underneath.
- Confident answers can be stale. In the demo, a sales agent answered cleanly from data that had stopped updating weeks earlier.
- Watch six signals. Retry rate, eval scores, trace depth, token use, first-seen resource access and customer tickets.
Most of an AI agent's life happens after launch, but most teams still focus on building and testing. That was the argument in the Data Science Connect webinar "Trusting AI in Production: Building Observability for AI Agents at Scale," a joint presentation by Monte Carlo and Snowflake.
The presenters were Sam Mittal, VP of AI Solutions and Global Director of ISV Applications, AI Startups & SaaS at Snowflake, and Ethan Post, Principal Forward Deployed Engineer, Agent Observability at Monte Carlo. Jenna Holland of Data Science Connect hosted.
Why do AI agents fail after they reach production?
Mittal said "production AI that works on Monday can silently degrade by Thursday." A proof of concept works in a demo because it doesn't need governance, fresh data or monitoring, he said. Production does. The common thread in failures is a disconnect between the data layer and the AI layer: the model doesn't know the data changed, the team doesn't know the model is drifting, and nobody notices until a customer complains.
Post cited a Monte Carlo study of more than 300 data and AI executives. Roughly a third of engineering teams said they were confident they had visibility into what their agents were doing, and 47% had found agents accessing data they shouldn't. Users discovered agent issues more than half the time, manual checks caught about 40%, and only 8% were caught proactively. "Production-grade applications need to handle production-grade variability," Post said.
How does Snowflake build trust into Cortex agents?
Mittal described six design principles for Snowflake Cortex Agents. The first is data access through role-based controls: "If your role can't see a column, neither can your agent," he said. The others are accurate retrieval, with Cortex Search for documents and Cortex Analyst for text-to-SQL; orchestration and memory across turns; tools such as web search and code generation; MCP interfaces in both directions; and trust, observability and governance.
That last layer includes native AI observability that logs every agent span, tool call and token count, Cortex Guard for runtime content safety, and evaluations through TruLens, which Mittal said Snowflake acquired. The hard part, he said, is operating agents at scale after they ship.
What should AI agent observability cover?
Post described what he called the anatomy of a trust framework, with four layers to check when an agent's answer looks wrong:
- Output: Is the response grounded and fit for purpose? Post showed groundedness scores broken down by workflow to pinpoint which one slipped.
- Behavior: Did the agent call the right tools and models in the right order? His example was a SQL tool call retried twice after it had already succeeded, adding latency and tokens.
- Performance: Error rates, time to first token and token use. A spike in tokens might trace back to a new agent version that retrieves the top 25 results instead of the top five.
- Data context: Is the underlying data accurate and fresh, and did the agent touch only the data on its allow list? "If you don't have good data, you're never going to have good agents," Post said.
Mittal noted that Snowflake's resource monitors track credits at the warehouse level, so a rising Cortex bill doesn't show which agent caused it. He called Snowflake's access controls prevention and Monte Carlo's monitoring detection.
What did the live demo show?
Post asked a Snowflake Intelligence sales agent for the top five branches over the past month and got a clean, confident answer in seconds. When he asked whether any branches had stale data, the agent found that none had received new data since mid-April.
In Monte Carlo, which reads the telemetry Snowflake writes to native tables without an SDK or moving data out of the account, he showed the trace of that conversation and the upstream invoice pipeline, which had raised four alerts in the past week. Monte Carlo had flagged the freshness problem when the feed stopped on April 13, he said. Post said Monte Carlo typically detects such issues within five to 10 minutes.
He also showed Monte Carlo's agents proposing baseline monitors and evaluations, mapping the agent's usual tool-call path, and troubleshooting a production case where a prompt adherence score fell from five to one. The cause was intermittent errors in one tool call tied to an upstream LLM call.
Which early warning signals should teams track?
Post listed six signals to watch, with or without Monte Carlo:
- Retry rate: steps that keep retrying when they shouldn't.
- Eval scores: sudden drops in output quality.
- Trace step blowout: rising call depth per execution.
- Token use: consumption that departs from forecasts.
- First-seen resource access: an agent fetching data it hasn't touched before.
- Customer tickets: user feedback linked back to agent behavior.
On whether to switch models, Mittal said to anchor on the KPIs set at the start and the drift you can tolerate. Post said teams should test model changes against ground-truth evaluations before switching, and keep monitoring afterward because testing won't catch everything. "The teams that are moving the fastest aren't taking more risk," he said. "They're the ones with tighter feedback loops."
See more on AI in production
Data Science Connect's 2027 schedule includes 24 free live webinars across the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Registration is coming soon; see the webinar schedule for upcoming sessions.
Frequently asked questions
What is AI agent observability?
Monitoring what agents do in production: output quality, tool calls and behavior, performance and token use, and the health of the data that feeds them, so problems are caught before users report them.
Why do AI agents fail in production when the demo worked?
Mittal said proofs of concept skip governance, data freshness and monitoring, and failures stem from a disconnect between the data layer and the AI layer.
What early warning signals indicate AI agent drift?
Post listed rising retry rates, falling eval scores, growing trace depth, unexpected token use, first-seen resource access and customer tickets.
Should teams switch to the newest LLM?
Only after testing. Mittal said to anchor on the original KPIs and acceptable drift, and Post recommended testing against ground-truth evaluations, then monitoring after the switch.
Watch: Trusting AI in Production: Observability for AI Agents with Monte Carlo & Snowflake | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
